About “Deprecation of the volatile flag”

1

Hi,
The 0.4 version deprecated the keyword volatile. I don’t understand the official document thoroughly. Can you explain the relationship among volatile, requires_grad, and with torch.no_grad()?In addition, how to modify the following two pieces of code when volatile=true or false. Thank!

def make_variable(tensor, volatile=False):
    """Convert Tensor to Variable."""
    if torch.cuda.is_available():
        tensor = tensor.cuda()
    return Variable(tensor, volatile=volatile)

    # evaluate network
    for (images, labels) in data_loader:
        images = make_variable(images, volatile=True)
        labels = make_variable(labels).squeeze_()

        preds = classifier(encoder(images))
        loss += criterion(preds, labels).item()

        pred_cls = preds.data.max(1)[1]
        acc += pred_cls.eq(labels.data).cpu().sum()

    loss /= len(data_loader)
    acc /= len(data_loader.dataset)

    print("Avg Loss = {}, Avg Accuracy = {:2%}".format(loss, acc))
2

Instead defining the volatile flag in the Variable you are now wrapping code blocks into a with torch.no_grad() block.

Your code snippet would thus change to:

with torch.no_grad():
    # evaluate network
    for (images, labels) in data_loader:
        labels = labels.squeeze_()

        preds = classifier(encoder(images))
        loss += criterion(preds, labels).item()

        pred_cls = preds.max(1)[1]
        acc += pred_cls.eq(labels).cpu().sum()

Also, don’t use the .data attribute anymore, as it might have unwanted side effects.

3

Thanks for your reply. Can volatile=False be the default value and can be removed directly? Or can the entire first paragraph of code be removed?

4

You can remove the usage of Variables completely.
The equivalent to volatile=True is the with torch.no_grad() block, so you should use it if you want to save memory and don’t want to calculate gradients for these operations.

5

So how to modify it when encountering volatile=False or requires_grad ?

6

I’ve posted a code snippet in my previous post.
Isn’t this working for you?

7

It can work. Thank you very much. What I asked above is an additional question.

8

In that case I misunderstood the question.
For volatile=False you can just use plain tensors without using the no_grad guard.